This is a very interesting post that fits nicetly with many other interesting statements Mike Pall made (such as his reference to the LuaJIT2 interpreter being faster than the LuaJIT1 jit compiler on some cases [http://lambda-the-ultimate.org/node/3851#comment-57646].)
On a related note, a similar problem is highlighted in a paper by Anton Ertl and David Gregg ([1]), where they claim that the slowdowns of an efficient interpreter to an optimizing native code compiler is 1:10, whereas the slowdown between an efficient interpreter and an ineffcient one is 1:100. Consequently, there seems to be a lot of optimization potential (though I admit that depending on the kind of the interpreter, it might be unlikely to stay within a slowdown of 1:10)
On a related note, a similar problem is highlighted in a paper by Anton Ertl and David Gregg ([1]), where they claim that the slowdowns of an efficient interpreter to an optimizing native code compiler is 1:10, whereas the slowdown between an efficient interpreter and an ineffcient one is 1:100. Consequently, there seems to be a lot of optimization potential (though I admit that depending on the kind of the interpreter, it might be unlikely to stay within a slowdown of 1:10)
[1]: Ertl, Gregg. "The Structure and Performance of Efficient Interpreters", 2001. (http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.69....)